Gumekn/Text2Trajectory__An-Agent-for-Corner-Case-Generation-in-Autonomous-Driving-Simulation

Text2Trajectory Agent is an LLM-driven tool for generating corner-case trajectories for autonomous-driving simulation. It accepts natural-language test requests, extracts parameters, autonomously plans generation strategies, mutates trajectories, and produces high-collision test data (target collision rates 30–50%) while keeping physical plausibility (>95%). Who uses it - AV testers, simulation engineers, QA/safety teams and researchers needing realistic, diverse corner cases for validation. Key features - LLM-based NLU (Claude/GPT-4) with prompt engineering and few-shot examples (parameter extraction >90%) - Autonomous strategy planning (normal/aggressive/dangerous/mixed) and scene matching - Trajectory engine: delta-a grid sampling, two-step probability selection, vehicle dynamics modeling - Collision detection (SAT algorithm), physical-constraint validation, automatic quality assessment and optimization suggestions - Multi-modal interfaces: CLI, Feishu bot, REST API (planned) - Data: Waymo loader, SQLite storage; visualization: 2D animation + MP4 export - Performance: e.g., ~50 traces in 45–60s, diversity pruning and continuous-learning pipeline (planned)

AI-related keywords found in this repo's description, topics, or README summary — grouped by category. Each badge links to the corresponding ranking detail page.